The use of machine learning and deep learning techniques to assess proprioceptive impairments of the upper limb after stroke.

The classifiers were trained to know a diagnosis that was already established by clinical and imaging assessment. The arm position matching task requires both hemispheres, so a stroke lesion may affect it more than some other proprioceptive tests. People with hemispatial neglect can perform the task poorly, and the authors did not give the models information about neglect. Single robotic parameters overlapped between stroke and control participants, so one parameter alone is hard to interpret. The machine learning models performed only slightly better than the overall task score. The authors note that machine learning methods require large datasets and many hours of data collection. The paper says bias in training data can carry into testing.

Funded by an Ontario Research Foundation Research Excellence grant, CIHR, and a Heart and Stroke Foundation of Canada grant. The supplied text lists funding but no competing interest declaration.

Do not read this as evidence that machine learning can diagnose stroke or guide treatment. The participants already had a clinical stroke diagnosis, and the authors say using kinematics to diagnose stroke is not practical.